Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: train ChatML datasets whose roles are capitalized (#15599)

Nilayf787c47c0c51
diff --git a/studio/backend/tests/test_chatml_role_names.py b/studio/backend/tests/test_chatml_role_names.py
new file mode 100644
index 00000000000..77254302a38
--- /dev/null
+++ b/studio/backend/tests/test_chatml_role_names.py
@@ -0,0 +1,95 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import pytest
+from datasets import Dataset
+
+from utils.datasets import format_and_template_dataset
+
+
+class _Tokenizer:
+ chat_template = "chatml"
+ eos_token = "</s>"
+
+ def apply_chat_template(
+ self,
+ messages,
+ tokenize = False,
+ add_generation_prompt = False,
+ ):
+ return "".join(
+ f"<|im_start|>{m['role']}\n{m['content']}<|im_end|>\n"
+ for m in messages
+ if m["role"] in ("system", "user", "assistant", "tool")
+ )
+
+
+def _conversation(system, user, assistant):
+ return [
+ {"role": system, "content": "Be brief."},
+ {"role": user, "content": "What is 2+2?"},
+ {"role": assistant, "content": "4"},
+ ]
+
+
+def _render(rows, column, format_type):
+ result = format_and_template_dataset(
+ Dataset.from_dict({column: rows}),
+ model_name = "Qwen3ForCausalLM",
+ tokenizer = _Tokenizer(),
+ format_type = format_type,
+ batch_size = 1,
+ num_proc = 1,
+ )
+ assert result["success"] is True
+ return list(result["dataset"]["text"])
+
+
[email protected]("column", ["messages", "conversations"])
[email protected]("format_type", ["auto", "chatml"])
+ "roles", [("System", "User", "Assistant"), (" system ", "USER ", " Assistant")]
+)
+def test_chatml_role_names_reach_the_template_in_lowercase(column, format_type, roles):
+ expected = _render([_conversation("system", "user", "assistant")], column, format_type)
+
+ assert _render([_conversation(*roles)], column, format_type) == expected
+ assert "<|im_start|>assistant\n4" in expected[0]
+
+
[email protected]("mapping", [None, {"prompt": "user", "completion": "assistant"}])
+def test_mapped_message_columns_reach_the_template_in_lowercase(mapping):
+ def render(user, assistant):
+ result = format_and_template_dataset(
+ Dataset.from_dict(
+ {
+ "prompt": [[{"role": user, "content": "What is 2+2?"}]],
+ "completion": [[{"role": assistant, "content": "4"}]],
+ }
+ ),
+ model_name = "Qwen3ForCausalLM",
+ tokenizer = _Tokenizer(),
+ format_type = "chatml" if mapping else "auto",
+ custom_format_mapping = mapping,
+ batch_size = 1,
+ num_proc = 1,
+ )
+ assert result["success"] is True
+ return list(result["dataset"]["text"])
+
+ expected = render("user", "assistant")
+ assert render("User", " Assistant ") == expected
+ assert "<|im_start|>assistant\n4" in expected[0]
+
+
+def test_lowercase_roles_skip_the_extra_pass():
+ from utils.datasets.format_conversion import normalize_chatml_roles
+
+ lower = Dataset.from_dict({"messages": [_conversation("system", "user", "assistant")]})
+ assert normalize_chatml_roles(lower, "messages") is lower
+ upper = Dataset.from_dict({"messages": [_conversation("System", "User", "Assistant")]})
+ assert [m["role"] for m in normalize_chatml_roles(upper, "messages")["messages"][0]] == [
+ "system",
+ "user",
+ "assistant",
+ ]
diff --git a/studio/backend/utils/datasets/dataset_utils.py b/studio/backend/utils/datasets/dataset_utils.py
index 7a94270b8c5..038f9875e87 100644
--- a/studio/backend/utils/datasets/dataset_utils.py
+++ b/studio/backend/utils/datasets/dataset_utils.py
@@ -18,6 +18,7 @@ from .format_conversion import (
convert_to_vlm_format,
convert_llava_to_vlm_format,
convert_sharegpt_with_images_to_vlm_format,
+ normalize_chatml_roles,
)
from .chat_templates import (
_custom_prompt_template_error,
@@ -399,6 +400,10 @@ def format_dataset(dataset, *args, **kwargs):
"requires_manual_mapping": False,
"warnings": [],
}
+ if info["final_format"] in ("chatml_messages", "chatml_conversations") and info["chat_column"]:
+ info["dataset"] = normalize_chatml_roles(
+ info["dataset"], info["chat_column"], bound.arguments["batch_size"]
+ )
return info
diff --git a/studio/backend/utils/datasets/format_conversion.py b/studio/backend/utils/datasets/format_conversion.py
index 8f57fc68f5c..7e2f2511332 100644
--- a/studio/backend/utils/datasets/format_conversion.py
+++ b/studio/backend/utils/datasets/format_conversion.py
@@ -214,6 +214,51 @@ def standardize_chat_format(
return result
+_CHATML_ROLES = {"system", "user", "assistant", "tool"}
+
+
+def normalize_chatml_roles(
+ dataset,
+ chat_column,
+ batch_size = 1000,
+):
+ def _fold_message(message):
+ if not isinstance(message, dict):
+ return message
+ role = _normalize_role_alias(message.get("role"))
+ if role in _CHATML_ROLES and message.get("role") != role:
+ return {**message, "role": role}
+ return message
+
+ try:
+ import pyarrow.compute as pc
+ roles = pc.unique(
+ pc.struct_field(pc.list_flatten(dataset.data.column(chat_column)), "role")
+ )
+ if not any(
+ _normalize_role_alias(r) in _CHATML_ROLES and _normalize_role_alias(r) != r
+ for r in roles.to_pylist()
+ ):
+ return dataset
+ except Exception:
+ pass
+
+ def _fold(examples):
+ return {
+ chat_column: [
+ [_fold_message(message) for message in convo] if isinstance(convo, list) else convo
+ for convo in examples[chat_column]
+ ]
+ }
+
+ return dataset.map(
+ _fold,
+ batched = True,
+ batch_size = batch_size,
+ features = getattr(dataset, "features", None),
+ )
+
+
def _content_text(content):
if isinstance(content, list):
return "\n".join(